@weaveintel/routing

Picks the right model for each request using policy, live health tracking, and weighted scoring — with an explainable decision.

Why it exists

You often have several models that could answer a request — a cheap fast one, a big careful one, a vision one — and choosing by hand for every call doesn't scale. Think of it like a hospital triage nurse: they read what's in front of them, know which specialists are on shift and which are overloaded, and send you to the right one with a reason. routing is that nurse for models: it reads the request, filters out models that can't do the job, scores the rest on cost and quality, avoids ones that are currently failing, and tells you why it chose what it chose.

When to reach for it

Reach for it when more than one model is a candidate and the choice should be automatic and auditable — cost ceilings, capability/modality constraints, fallback chains, health-aware avoidance of a flaky provider. If you always call the same single model, you don't need a router. The routing contracts (ModelRouter, RoutingPolicy, RoutingDecision) live in @weaveintel/core; this package is the implementation.

How to use it

import { SmartModelRouter, ModelHealthTracker, ModelScorer } from '@weaveintel/routing';

const router = new SmartModelRouter({
  scorer: new ModelScorer(),
  health: new ModelHealthTracker(),
});

const decision = await router.route(
  { taskType: 'summarize', maxCostPer1kTokens: 0.5 },
  ['claude-model', 'fast-cheap-model', 'vision-model'],
);

console.log(decision.selectedModel, decision.reason);

What's in the box

• Router — SmartModelRouter (the top-level decision maker).
• Health — ModelHealthTracker (records successes/failures to steer away from failing models).
• Scoring — ModelScorer weighs cost vs. quality vs. capability.
• Policy helpers — filterByConstraints, filterByCapability, filterByModality, filterByCostCeiling, roundRobinSelect, fallbackChainCandidates.
• Task inference — inferTaskType guesses the task from the request.
• Capability flags — getModelCapabilityFlags, getModelContextWindowK, getModelMaxOutputK.
• Runtime slot & seeds — createRuntimeRoutingAdapter, plus DEFAULT_ROUTING_POLICIES, DEFAULT_MODEL_PRICING, DEFAULT_TASK_TYPES.

Single entry point: @weaveintel/routing.

License

MIT.